China Pivots AI Strategy From Foundation Models to Autonomous Agents

China is pivoting its artificial intelligence strategy away from the development of large-scale foundation models in favor of the creation and deployment of AI agents. This shift marks a transition from passive content generation to autonomous task execution, signaling a move toward AI that can operate independently to achieve specific goals.

According to a state report detailed by Bloomberg, this strategic realignment is expected to fundamentally alter the nation's computing infrastructure. AI agents are projected to drive nearly tenfold annual growth in China's computing demand over the next two to three years. This surge is primarily driven by the transition from training-heavy workloads to inference-heavy operations. By 2029, inference computing is projected to account for 80% of China's computing-power market, officially surpassing the demand for training-related computing.

The End of the Model War

For several years, the global AI race was defined by a "model war," where developers in China and abroad competed to increase parameter counts and amass larger training datasets for Large Language Models (LLMs). However, the industry has reached a point where the marginal utility of simply building larger models is plateauing while the financial and energy costs of training continue to skyrocket. In response, the focus has shifted toward "agentic AI"—systems capable of using external tools, planning multi-step actions, and executing complex workflows without constant human intervention.

From Chatbots to Digital Workforce

This pivot signals a critical transition from viewing AI as a sophisticated chatbot to treating it as a functional workforce. By prioritizing agents and inference, China is emphasizing the practical economic utility of AI within industrial sectors and government administration. The goal is to move beyond generative text and images toward systems that can autonomously manage logistics, coding, or administrative governance.

Infrastructure Implications

The massive projected increase in inference demand creates an immediate and critical need for specialized hardware and energy infrastructure. Unlike training, which happens in concentrated bursts on massive clusters, millions of active agents running in real-time require a distributed and highly efficient inference layer. This shift will likely dictate future investment in semiconductor procurement and data center architecture as the country prepares for a landscape dominated by autonomous agents rather than static models.

Future Outlook

As the industry moves toward this agent-centric model, the primary metric of success is shifting from benchmark scores to real-world task completion rates. Observers will now be watching for the deployment of these agents in large-scale industrial pilots and the ability of China's hardware supply chain to meet the looming 80% inference market share projection.